Learning Robust Diachronic Representations of Ancient Greek Letterforms

Analyzing ancient texts presents a significant challenge due to the vast temporal variations in handwriting styles over centuries. This paper tackles this problem by focusing on ancient Greek, one of the longest continuously used writing systems, to develop robust diachronic representation learning. The researchers introduce three new datasets: Hell-Char for training (3rd-1st centuries BCE) and two evaluation sets, PaLit-Char (2nd-5th c. CE) and Med-Char (9th-14th c. CE), to capture this temporal evolution. To overcome issues like symbolic variation, scarce data, and systematic degradation common in historical manuscripts, two key strategies are proposed. First, a similarity-weighted supervised contrastive loss is used to bias embeddings, dynamically estimating inter-class similarities to improve character separation. Second, a lacuna-driven augmentation scheme simulates realistic manuscript corruptions, making the models more resilient to noise. When trained with these innovative strategies, both a lightweight Convolutional Neural Network (CNN) and a pre-trained ResNet demonstrated strong recognition performance. They produced embeddings that more effectively separated character classes compared to traditional methods like PCA or generic pre-trained models. These interpretable embeddings facilitate tasks such as clustering, identifying stylistic subgroups, and visualizing the diachronic evolution of letterforms, offering a transferable paradigm for representation learning under challenging historical data conditions.